📚 CLI tool to fetch llms.txt files for your npm dependencies

Install it globally

npm install -g get-llms

or just run it with npx

npx get-llms

Fetch llms.txt files for all dependencies in your package.json:

get-llms

You can fetch specific packages without needing a package.json:

get-llms zod react-native

Specify a custom path to package.json (default: ./package.json)

get-llms --package ./packages/core/package.json

Filter which dependency types to include. Comma-separated list of:

• prod - dependencies
• dev - devDependencies
• peer - peerDependencies
• optional - optionalDependencies
• all - all dependencies (default)

get-llms --deps prod,dev        # Only production and dev dependencies
get-llms --deps prod            # Only production dependencies

Specify packages directly to fetch:

get-llms react @types/react

Custom output directory (default: docs/llms)

get-llms --output ./context/dependencies

Filename pattern for output files. Use {name} as a placeholder for the package name (default: {name})

get-llms --filename "llms-{name}"

File extension for output files (default: txt)

get-llms --extension md

Preview what would be done without writing any files

get-llms --dry-run

Strategy when llms.txt is not found:

• none (default) - Skip packages without llms.txt
• readme - Fall back to README.md from GitHub
• empty - Create an empty file with a placeholder message
• skip - Same as none

get-llms --fallback readme     # Use README.md as fallback
get-llms --fallback empty      # Create empty files

With readme fallback:

get-llms --fallback readme

Results in:

✅ package-name: Using readme fallback -> package-name.txt

With empty fallback:

get-llms --fallback empty

Creates files containing:

# package-name

No llms.txt found for this package.

Only show errors (minimal output)

get-llms --quiet

Show detailed output including:

• Network requests being made
• URLs being checked
• Debug information

get-llms --verbose

Note: If both -q and -v are specified, the last one wins.

Customize how special characters in package names are sanitized:

Character to replace spaces in filenames (default: _)

get-llms --space-replace "_"

Character to replace slashes in filenames (default: -)

get-llms --slash-replace "-"

Character to replace @ in scoped package names (default: empty string)

get-llms --at-replace ""

Example: @types/node becomes types-node.txt with default settings

get-llms \
  --fallback readme \
  --extension md \
  --output ./docs/handbook

get-llms \
  --deps prod \
  --filename "{name}-reference" \
  --extension txt \
  --output ./context/prod

get-llms @types/node @types/react \
  --space-replace "_" \
  --slash-replace "_" \
  --at-replace "at_"

This might create:

• at_types_node.txt
• at_types_react.txt

get-llms \
  --deps all \
  --dry-run \
  --verbose

get-llms \
  --deps prod \
  --fallback empty \
  --output ./docs/llms-production \
  --quiet

By default, files are saved to:

docs/llms/
  ├── zod.txt
  ├── react-native.txt
  └── [package-name].txt

With --fallback readme, successful files may include fallback indicators in the summary.

The tool searches for llms.txt files in this order:

1. Check the package's package.json for an llms or llmsFull field
2. Try standard URLs: {homepage}/llms.txt and {homepage}/docs/llms.txt
3. If GitHub repository, search README for links containing "docs"
4. Apply fallback strategy if specified

flowchart TD
    A([package]) --> B{is there an 'llms' key in the 'package.json'?}

    B -- yes --> FOUND([we've found it])
    B -- no --> C{is the package homepage a github link?}

    C -- yes --> D{does the github 'readme.txt' mention the word 'docs' in a hyperlink?}
    C -- no --> E{does 'link/llms.txt' return a txt file?}

    D -- yes --> E
    D -- no --> NOFILE([there likely isn't a 'llms.txt' for that package Either use the 'README.txt', or use an external service like context7])

    E -- yes --> FOUND
    E -- no --> F{does 'link/docs/llms.txt' return a txt file?}

    F -- yes --> FOUND
    F -- no --> NOFILE

• 0 - Success
• 1 - Package.json not found or cannot be parsed

Issues and contributions welcome! Report bugs at: https://github.com/balazshevesi/get-llms/issues, or contact me on twitter @balazs_hevesi

The llms.txt spec is still young and un-opinionated. Some packages and some documentation-sites already ship an llms.txt file, others don’t, and the ones that do sometimes invent their own markdown dialect.

As adoption grows we expect the format to stabilise, turning “antigenic-coding” into a first-class workflow.

Today the cleanest way to publish documentation with an llms.txt is to add an "llms" key to package.json:

"llms": "./README.txt"

When the registry starts indexing that field we’ll be able to fetch the exact documentation that and feed it into the coding-agent's context. (this will also require the URLs for fetching documentation versions to be standardized)

Example: zod

Because the format is still fluid, treat llms.txt as best-effort documentation:

• 90 % of the time an agent can slurp the file directly.
• If it’s 25k tokens or more, you might want to fall back to a cache, a RAG layer, or a service such as Context7. (AI-agents such have a limit on the file sizes they're willing to read, and will truncate the file to fit their context window)

Vibe-coded apps are about to become legacy and enter a maintenance phase. When that happens, ai coding-agents will need the docs for old versions of the packages. llms.txt is the lowest-friction way to provide the ai with the correct documentation.

External services (like Context7) are great, but bundling the required context is simpler, cheaper.

MIT License